Mergers, Joint Ventures or Investments as Routes to Controlling AI Model Costs

Just how artificial intelligence model providers might improve their economics is a key business model issue. 

A shift to inference operations also emphasizes the importance of reducing cost per token at scale. 

source: McKinsey 

Where software often has marginal costs close to zero, use of AI models seem to have costs that scale almost linearly with usage, so marginal costs are high. 

A few key business issues are clear enough.

A greater shift of cost towards fixed cost rather than variable cost seems necessary. For example, model creators could move in the direction of owning their compute infrastructure rather than renting cloud capacity.

The problems with high variable cost are clear:

• Unit economics

• Cash burn and need for capital injections

• Competitive pressure

• Capital allocation

Right now, model queries do not show software-style marginal cost trends, where marginal costs are close to zero. 

Every additional user or query drives proportional costs for GPUs, power, and data center capacity. 

Reports suggest inference alone consumes 50 percent of revenue in some cases. Some suggest the problem is worse, with at least some model providers spending more than 100 percent of revenue on compute services. 

By itself, that might not be existential, as model providers are in the early stages of growth, meaning incremental revenues would not be expected to cover the full costs of creating and operating the models at scale.  

But observers do worry about marginal costs that seemingly do not have software-style economics with near-zero marginal costs.

Cash burn and capital intensity also are issues. Rapid revenue growth is offset by even faster cost scaling, which necessitates high investor funding requirements, borrowing, equity raises. 

Pricing issues also are an issue. When model suppliers raise prices, introduce usage-based tiers, or limit free access, they risk customer churn or slower adoption, even as they address revenue issues. 

Strategic vulnerabilities also exist when model suppliers are dependent on external suppliers for crucial computing services (operating costs, ability to manage surges of demand). 

Capital allocation is an issue as well. One might argue that model builders should divert capital into compute infrastructure, but that is hugely expensive and detracts from the job of developing the next generation of models.

So the issue is how to fix the problem. Revenue growth with scale obviously helps, but doesn't solve the marginal cost issue. Efficiency improvements, owned infrastructure and pricing innovations all will play a role.

Smaller or distilled models, sparse activations, better architectures, prompt caching, batching, quantization, and routing to cheaper models for simpler tasks will happen.

Inference costs per token have dropped significantly in some cases, allowing gross compute margins to  improve.

source: Beth Kindig 

Stranded assets always are a problem, so higher GPU utilization rates help. So do custom silicon and algorithmic advances.

Owned infrastructure is a partial answer. Model builders and compute suppliers are investing heavily in their own data centers and chips (Anthropic's $50 billion commitments for custom U.S. facilities with partners like Fluidstack; OpenAI's Stargate).

Revenue models also are adjusting. Higher enterprise pricing, usage-based tiers, value-based pricing (charge relative to delivered value, not just tokens) and premium features or apps are introduced. 

Given all that, one logical historical precedent is for mergers and acquisitions that place model building and compute functions under a single ownership. On the other hand, antitrust regulations will probably tend to restrict options for some of the most-likely buyers (Alphabet, AWS, Microsoft, SpaceX, for example).

So other forms of cooperation are likely to develop. 

Expect partnerships, joint ventures, dedicated capacity deals, and partial ownership rather than full-scale mergers and acquisitions, which will face antitrust opposition. 

When feasible, model builders are creating their own compute infrastructure. OpenAI's Stargate project with Oracle and SoftBank (up to $500B, multi-gigawatt scale) provides an example. Investments in neocloud suppliers is another example. 

But that might be the exception to the rule. Anthropic, for example, has chosen to sign big supply deals rather than build its own facilities.

Partnerships for Tensor Processing Units, Trainium, and other accelerators reduce reliance on expensive third-party GPUs and improve efficiency also are growing. 

But full vertical integration might not be the immediate or mid-term path forward, partly for regulatory scrutiny reasons; partly for capital intensity reasons and partly for business diversity reasons. Both model builders and compute infra providers prefer a diversity of partners. 

So joint ventures and consortia that also have the advantage of off-balance-sheet implications will happen. 

In summary, tight strategic integration and partial ownership rather than blockbuster mergers are the main approaches. That avoids regulatory opposition and also is capital efficient. 

Solving AI Model Marginal Cost Issues

Profit margins arguably are the key business issue for frontier artificial intelligence model providers.

Where software businesses have tended to have low marginal costs, with high profit margins at scale, AI businesses tend to have no such advantage. 

Inference and training have heavily variable costs tied to usage and scale. In other words, cost per query scales mostly in linear fashion. 

That is quite unlike traditional software, with 80 percent to 90 percent gross margins at scale, since marginal cost is quite low. 

And that might have important ramifications for pricing models.

Traditional software can often support very high gross margins because the marginal cost of serving one more user is close to zero, while AI products can have materially higher variable costs because each additional query or token consumes compute.

That difference pushes AI businesses toward pricing and operating models that track usage, outcomes, or hybrid bundles rather than pure seat-based subscriptions.

For classic software, once the product is built, distributing another copy is cheap, so revenue can scale much faster than cost. 

In AI, each extra interaction may add inference, context, storage, and orchestration costs, so the economics can look more like a utility or metered service than a pure software license.

That does not mean AI cannot be highly profitable; it means profitability depends more on model efficiency, pricing design, and control over the inference stack.

The emerging pattern is hybrid monetization mixing flat rate charges and usage:

• subscriptions for baseline access

• usage-based credits

• metered throughput

• outcome-based pricing for heavier AI usage.

If additional usage carries significant marginal cost, sustainable AI business models usually combine four elements: 

• cost-aware product design (cost per interaction)

• pricing aligned to actual usage intensity (tiered pricing)

• infrastructure leverage (own rather than rent)

• AI feature packaging (proprietary data, workflow integration). 

A practical way to think about it is this: the business must make sure marginal revenue per interaction stays above marginal cost per interaction, with enough spread to fund sales, support, R&D, and model improvement. 

That often means charging more for complex tasks, throttling or downgrading expensive requests, or bundling some usage into plans while monetizing power users separately.

Issue	Why it matters	Possible solution
High per-use compute cost	Margins can compress as usage rises	Tiered pricing, credits, or metered billing tied to task intensity
Third-party model dependence	Vendor price changes and rate limits can hit gross margin	Own more of the inference stack, add model abstraction, or use open-source/smaller models where feasible
Weak price predictability	Customers dislike uncertain AI bills, slowing adoption	Pre-commits, usage caps, transparent dashboards, and clear credit buckets
Flat seat pricing mismatch	Power users can consume far more than average users	Hybrid subscription plus usage-based overages or action-based pricing
Hard-to-measure outcomes	Outcome pricing is difficult to verify and negotiate	Start with activity-based metrics, move to outcomes only where measurable and automated
Scaling costs faster than revenue	Growth can worsen economics instead of improving them	Cost monitoring, workload routing, batch inference, and feature-level margin analysis
Commoditization risk	Models alone are easy to copy	Build proprietary data, vertical specialization, and workflow integration

AI winners are more likely to be firms that control costs, can manage usage and price in a way that reflects both value delivered and the true marginal cost of serving that value.

Anthropic Fable 5: a Marketing Platform Gone Wrong

Has Anthropic poisoned its own well by arguing its new Mythos model was so dangerous it couldn’t be released publicly?

After Amazon researchers discovered a jailbreak of Fable 5 (Mythos with guardrails), the U.S. government banned use of both Mythos and Fable 5, exactly the kind of government scrutiny that could disrupt Anthropic’s  business most.

Amazon CEO Andy Jassy was reportedly the source of security concerns that led the U.S. government to force Anthropic to shut down Fable 5 and Mythos 5 for every customer, also imposing an export control ban on both models.

Oddly enough, Amazon is one of Anthropic’s largest investors, having put in billions and receiving a $100 billion cloud spending commitment in return. 

Anthropic says the jailbreak technique surfaced “a small number of previously known, minor vulnerabilities.” 

It called the government’s response disproportionate and said the capabilities causing concern are already available in other publicly accessible models. The shutdown affected every customer globally because Anthropic cannot filter foreign nationals from U.S. users in real time.

For Anthropic, the immediate question is how quickly it can restore access. And that will happen.

For the broader industry, the precedent is what matters. Anthropic has been lobbying for more regulatory power over AI. It just found out it might get what it wants. 

Anthropic's own CEO publicly called for the government to have authority to block dangerous model releases. The government then used such calls against Anthropic two days later.

Optimists might argue that government restrictions on dual-use technologies might slow down technology diffusion, but cannot halt it.

The general pattern across technology history is this: governments impose controls with genuine security rationale, those controls prove partially effective in the short term, generate substantial unintended costs, and are eventually relaxed or circumvented.

The regulated party frequently bears costs that accrue to unregulated competitors.

Technology	Regulation	Period	Short-Term Effect	Long-Term Outcome
Cryptography, Encryption (PGP, SSL)	U.S. export ban; treated as a munition under ITAR; Clipper Chip mandate attempt	1975 to 2000	Phil Zimmermann prosecuted 3 years for posting PGP online; RSA export restricted New America · Brookings	Controls collapsed when it became obvious software could not be bottled; Clinton-era deregulation followed. In subsequent decades, virtually all predictions about encryption's economic benefits were borne out — SSL, SSH, electronic banking, e-filing, and VPNs depended on exactly the strong encryption the government tried to restrict.
Cryptography	Export controls treated software as equivalent to munitions	1990s	The Zimmermann case dragged on three years before being dropped, and export control laws were eventually rewritten after it became obvious that software couldn't be contained like rocket motors. Crowe LLP · Wassenaar/arxiv	The EU abolished cryptographic export controls within the union in 2000, a decision subsequently adopted by the Clinton administration; more recent trends have involved further relaxation rather than tighter controls.
CoCom, Cold War Computing	Coordinating Committee for Multilateral Export Controls: Western embargo on technology exports to Soviet Bloc	1949–1994	In the period 1951–1967 CoCom performed reasonably well despite its known limits. After that, enforcement eroded. ChinaTalk/Ottinger · Texas Nat'l Security Review	CoCom's effectiveness was reduced throughout its history by overt non-compliance, differences between member nations, the secretive regime, and the financial incentives of tech exporters. CoCom did not prevent the USSR from accessing key technologies; the size of the control regime expanded over time, weakening enforcement and encouraging defection. Disbanded 1994.
Toshiba-Kongsberg Scandal	Sanctions on Toshiba for selling machine tools enabling quieter Soviet submarine propellers	1982–1987	Washington imposed sanctions on Toshiba worth $30 billion amid congressional claims the breach jeopardized U.S. national security. AFSA	The $30 billion estimate was primarily based on a hyperbolic claim that the entire U.S. nuclear submarine fleet would need to be replaced — the actual magnitude of the damage remained unclear. Sanctions were eventually relaxed; Toshiba survived.
Nuclear Technology	Nuclear Non-Proliferation Treaty (NPT); Nuclear Suppliers Group; export controls on enrichment technology	1968–present	By the 1950s, it became clear that pure denial had neither stopped the Soviet Union nor the UK from acquiring nuclear weapons. The 1954 revision of the Atomic Energy Act reflected a shift from prevention-through-denial to influence-through-cooperation. CNAS · AADS	U.S. dominance in commercial nuclear energy in the 1970s allowed the U.S. to keep Taiwan and South Korea from developing their own nuclear facilities — a genuine success. But Pakistan, India, North Korea, and Israel all acquired nuclear capability despite controls. NPT remains the most durable tech-control regime, though partial.
Semiconductors, AI Chips (China)	BIS October 2022 controls on NVIDIA A100/H100 and manufacturing equipment; tightened 2023, 2024, 2025	2022–present	Implementation significantly disrupted China's semiconductor ecosystem, causing price spikes for some device types and forcing workforce reductions. CSIS/HSToday · CSIS	Loopholes, alternative approaches, and unforeseen outcomes have diminished long-term efficacy — the very restrictions designed to hinder Beijing may instead be accelerating China's domestic progress. Companies impacted by the October 2022 controls, on average, even outperformed comparable unaffected companies, exhibiting higher increases in R&D spending and patent filings — a counterintuitive outcome driven by AI chip demand.
Semiconductor Controls	EAR controls tightened repeatedly; allied coordination with Netherlands, Japan	2022–present	According to research, versions of Intel's Xeon Gold and NVIDIA GeForce RTX chips can be bought on Taobao; students at Tsinghua report being able to "easily circumvent" restrictions on U.S. EDA software. Texas NSR	CoCom did not prevent the USSR from accessing key technologies; the current regime is similarly porous, and China is a more adept target. Long-term verdict still open.
AI Models Fable 5 / Mythos 5	U.S. Commerce Dept. emergency export control directive; all foreign nationals barred	June 2026	Anthropic had to cut off access to both models for all customers worldwide because it could not rapidly implement nationality-based filtering. Anthropic disputed the severity of the jailbreak, arguing many of the same vulnerabilities could be discovered using other publicly available models. Fortune · Axios · State of Surveillance	Too recent to assess long-term outcome. The next letter may go to a different vendor; the policy tools used here will be available to every future administration. Watch for restored access accompanied by new conditions such as nationality verification, tighter pre-deployment review, or federal observer access.

Several durable findings emerge from this body of evidence:

• Unilateral controls erode without multilateral coordination

• The control regime tends to expand and become self-defeating

• Denial often accelerates indigenous capability in the target

• Software and information resist containment

• The regulated party bears costs; competitors do not. 

Rival Sam Altman, of OpenAI, says Anthropic’s fear-based marketing now has come back to bite Anthropic. 

AI Model Pricing Eventually Moves to a Cloud Model

So far, language model pricing based on usage closely mirrors the early patterns of cloud computing, .

Which matured into a massive, lower-margin commodity market with heavy optimization practices.

AI inference is on a similar but possibly accelerated trajectory.

We might reasonably expect:

• Price deflation and commoditization

• More sophisticated pricing models

• Shift to agentic operations

• Hybrid platforms

• More cost governance measures.

source: Rohan Modi 

It is reasonable to expect prices for baseline capabilities will fall (potentially another 5-10x over years).

Frontier capabilities will remain premium but get cheaper over time, as well.

We probably should see "good enough" models for most tasks, mirroring how basic cloud virtual machines became very affordable.

Reserved-like plans, volume discounts and sustained use discounts already are emerging, with more pricing granularity. 

The shift from chatbots to agents will likely lead to pricing based on:

• Per-agent/subscription (like "digital employee" salaries)

• Outcomes (pay per resolved ticket, qualified lead, successful workflow)

• Hybrid models (base + usage + performance bonuses).

Overall, the shift is from pricing based on “raw compute” to “managed services” with value-based elements.

source: Rohan Modi 

Overall, expect AI pricing to become more predictable, value-aligned, and cheaper per capability unit, much as cloud computing evolved.